Leveraging Structured Knowledge to Enable Efficient AI Assistants
Author(s)
Samel, Karan Manoj
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Advances in Artificial Intelligence systems have evolved from classifying individual images or texts to recently providing fluid conversations and generating works of art based on a user’s criterion. These advances have leveraged the amount of data and computation power at scale developed over time. To generalize such systems to novel inputs or use cases, we expect them to operate over logical steps and involve knowledge it has learned along the way. This follows the idea that we as humans also leverage a core set of principles and facts to generalize to new conclusions. Therefore, when intelligent systems leverage prior curated heuristics and knowledge of the task at hand, they avoid the need to see all novel input examples during development. These principles allow us to develop these systems in a data as well as compute efficient manner.
In this thesis, we focus on developing the knowledge component of these systems. We demonstrate how to leverage different knowledge structures: knowledge graphs, hierarchies, and instructional steps to enable more efficient intelligent systems. For each knowledge source, we first identify methods to mine these structures from raw data. Then for each structure type, we introduce frameworks that leverage these structures to improve the efficiency of different downstream tasks to assist users. These knowledge sources can be generally applied across different types of tasks, from helping recommend e-commerce products to suggesting instructional videos given a user’s need. In each task, we demonstrate use cases where these external knowledge sources help and identify future directions for improvement.
Sponsor
Date
2024-12-03
Extent
Resource Type
Text
Resource Subtype
Dissertation